HiExplore scores any question 0–100 across six dimensions — verifiability, depth, scarcity, social value, innovation, feasibility. Below 60? Don't research it yet. Above? A team of AI agents breaks it into a question tree and explores it for days — every claim must carry a falsifiable hypothesis, and when the AI gets stuck, it asks real humans through your phone.
Most "this seems interesting" questions die on verifiability and feasibility. Finding out costs you 10 seconds instead of three weeks.
Unfalsifiable, answered a thousand times, no way to know you were right.
Checkable, specific, nobody has published it.
Drop in the question you actually want answered — or pick one from the question board.
QVS returns a 0–100 score with a radar breakdown and sharpening suggestions. Below 60, it tells you not to bother yet.
5–8 directions → specialized agents (researcher, analyst, reviewer…) → loop of decompose → execute → review → record, producing falsifiable hypotheses, notes, and simulations.
When the AI hits a question only reality can answer, it's pushed to real humans via a companion phone app. Decision timeline can replay any branch.
| Generic AI search / Deep Research | HiExplore | |
|---|---|---|
| Question value gate | ✗ answers everything | ✓ QVS 6-dim score, <60 = don't bother |
| Claim reliability | looks right | every node requires a falsifiable hypothesis; reasoning ≠ evidence |
| Simulation vs reality | blended | simulations explicitly marked "not evidence", with reality check + probe hint |
| Stuck? | keeps hallucinating | pushes the question to real humans / real data (phone app) |
| Your data | platform cloud | your browser / local Markdown, Obsidian-friendly |
| Models | locked in | bring your own: DeepSeek, Qwen, Claude, local Ollama |
| Code | closed | MIT, self-hostable |
Decompose a big question into verifiable sub-questions; an agent team grinds 24/7 while you make the calls.
Run every topic idea through QVS before investing; exploration notes become your content library.
Turn "things I want to understand" into long-running research projects with a personal knowledge graph.
Data lives in your browser or local Markdown (Obsidian-friendly). Export a note or the whole vault as a zip.
Fully local with Ollama (data never leaves your machine), or any OpenAI-compatible endpoint — DeepSeek, Qwen, Claude.
Hide your API key behind a serverless function so it never reaches the browser.
Register HTTP-API devices (sensors, robot arms) — agents read from and act on them mid-exploration, with an auditable call log.
No — try it without signing up. Login (WeChat or email) gives persistence and a higher daily quota.
By default it stays in your browser / local files. Only aggregate stats (active time, token usage) are synced to the cloud.
Any OpenAI-compatible endpoint: DeepSeek, Qwen, Claude, or fully local via Ollama/LM Studio. A desktop app is recommended for local models.
They're better at answering; HiExplore is about whether a question deserves answering, and keeping claims falsifiable and checked against reality. You can even run their reports through HiExplore as an audit.
Yes — MIT, self-hostable, no backend required. github.com/chenhaiyan123/ai-auto-explorer